Tencent to Acquire Majority Stake in Manus — AI Agent Reshuffle Follows Beijing's Block of Meta

The M&A drama surrounding AI agent startup Manus has taken a new turn. According to the Financial Times, Tencent is in talks to acquire a majority stake in Manus, with the deal reportedly valued at $2 billion — the same figure Meta had put on the table.

Beijing had previously blocked Meta's attempted acquisition of Manus (also a reported $2 billion deal) on national security grounds. Tencent reportedly sees strategic value in the deal, as Manus's technology overlaps with its own agent ambitions for WeChat. Meanwhile, U.S. venture capital firm Benchmark is not expected to take part in the transaction.

The pattern of Chinese AI agent companies shifting from acquisition targets of Western tech giants to 'domestic-led' ownership suggests the U.S.–China divide in AI is deepening further.

Databricks Benchmark: pi-coding-agent Delivers 2x Cost Efficiency, GLM 5.2 Closes In on Opus 4.8

A coding agent benchmark Databricks ran against its own multi-million-line codebase is drawing attention.

The biggest surprise: 'pi-coding-agent', reportedly used internally at Anthropic and designed around Bash with a minimal toolset, achieved up to 2x the cost efficiency of Claude Code (CC) and Codex while matching or beating them on pass rate. The finding that more tools doesn't necessarily mean better results challenges prevailing assumptions about agent design.

Also noteworthy: GLM 5.2 outperformed GPT 5.5 at both its high and xhigh settings, landing nearly on par with Opus 4.8 high. Databricks has experience training large LLMs of its own, and the community generally regards its analysis as reasonably credible.

This is likely to reinforce the growing view that, for coding tasks, the practical gap between expensive frontier models and Chinese-origin open models is steadily narrowing.

Meta's AI Image Detector Misses Its Own AI-Generated Images

According to Reuters, Meta's AI image detection tool has been found to fail to correctly identify some of Meta's own AI-generated images — particularly cropped ones.

If a simple edit like cropping an image can evade detection, it calls into question the effectiveness of transparency and labeling regimes for AI-generated content. The fact that a platform operator cannot even detect its own outputs underscores just how high the technical bar for AI content identification really is.

SpaceX Unveils AI Satellite 'Starmind'

SpaceX has published a page for an AI-equipped satellite named 'Starmind'. Details are limited, but SpaceX's stated direction of running AI processing on orbital platforms could open up a new frontier at the intersection of space infrastructure and AI.

Potential applications include low-latency inference combining satellite communications with AI, and real-time analysis of Earth observation data — though specific capabilities and a launch timeline remain unknown for now.

AI Cameras on Garbage Trucks — Waste Scanning Detects Code Violations and Sparks Privacy Debate

The city of Cape Coral, Florida, has begun fitting garbage trucks with AI cameras to scan residential areas for code violations. The system automatically detects local ordinance violations such as overgrown weeds and abandoned trash.

While it may look like efficient municipal monitoring on the surface, critics have raised concerns about resident privacy and over-surveillance. The debate over a society where 'garbage trucks scan your house every week' is a noteworthy new example of AI surveillance technology embedding itself into everyday life.

Experiment Getting LLMs to Think in Latent Space Draws Attention

Developer Oli-26 has published experimental results on GitHub in which an LLM is made to think directly in latent space. Instead of decoding to text, the model reasons in its internal representations — an approach that yields just the result without a human ever reading the model's 'thinking'.

The approach could lead to faster inference and lower token consumption, but it also carries the drawback of an opaque intermediate process. How far this method — essentially an 'invisible' version of Chain-of-Thought — can go toward practical use is something to watch in future research.

---